How can low-relevance chunks be filtered from Brave LLM Context?

Definition: Brave's context_threshold_mode parameter controls how aggressively LLM Context filters content for relevance before including it in the context returned to an AI agent. It accepts strict, balanced, lenient, or disabled; when it is unset, Brave applies its calibrated default threshold.

Key facts

• Brave describes LLM Context as web search with pre-extracted content optimized for AI agents, LLM grounding, and RAG pipelines in its LLM Context documentation.
• The documentation defines context_threshold_mode as the relevance threshold for including content and gives null as its default.
• strict uses a higher threshold and returns fewer but more relevant results. balanced is described as a balance between coverage and relevance.
• lenient uses a lower threshold and can include less relevant content, while disabled removes threshold filtering and returns all extracted content.
• Brave's LLM Context API reference describes the parameter as the mode used to determine the threshold for including content.

What does context_threshold_mode control?

Brave's documentation says that the parameter controls how aggressively the API filters content for relevance. The same page labels the setting a “Relevance threshold for including content.” The API reference presents it as a string enumeration, so the setting belongs in the request that retrieves the LLM context.

The setting concerns content selection. Brave describes LLM Context as returning extracted page chunks and source metadata in a shape and length designed for an LLM, while traditional Web Search returns titles and short snippets for a results page. The API reference calls the returned units “snippets (or chunks of text).”

What are the four modes?

Brave's threshold mode table expresses four different threshold levels:

• strict applies a higher threshold. The documentation pairs this with fewer but more relevant results.
• balanced provides a balance between coverage and relevance.
• lenient applies a lower threshold. The documentation says that it produces more results and may include less relevant content.
• disabled applies no threshold filtering and returns all extracted content.

The four values can be represented directly in a request parameter:

context_threshold_mode=strict

Replace strict with balanced, lenient, or disabled when the required relevance and coverage tradeoff changes. The documented behavior is about the threshold used for content inclusion, so it does not define a page ranking position or a citation outcome.

What happens when the parameter is omitted?

When context_threshold_mode is left unset, Brave's documentation says that the API applies its own calibrated default threshold. The documentation table lists the default as null, which is the request-level representation of an omitted setting.

Threshold filtering is separate from context-size controls. Brave's API reference gives 8192 as the default and 32768 as the maximum for approximate tokens in context. It gives 50 as the default and 256 as the maximum for snippets, described as chunks of text. The documentation also states that token budgets are the binding limit when they allow more snippets than the configured snippet counts.

The page's 2026-07-31 changelog says that maximum_number_of_snippets and maximum_number_of_snippets_per_url no longer constrain the response when the token budget allows more snippets. This means a threshold mode can determine which content qualifies for inclusion while the token budget limits the amount of context returned.

Related terms

• Brave Search API
• Does Brave use its own index for AI answers?
• Context window

FAQ

Which mode should you use for fewer low-relevance chunks?

Use strict when the intended behavior is a higher threshold with fewer but more relevant results, as described in Brave's threshold mode table.

What does balanced do?

balanced provides a balance between coverage and relevance, according to Brave's threshold mode table.

What happens with disabled?

disabled removes threshold filtering and returns all extracted content, subject to the context-size limits documented for the API.

How can you check the configured mode?

Inspect the request parameter and look for context_threshold_mode with one of Brave's four documented values. An omitted parameter corresponds to the documented null default.

Sources

1. Brave Search API, "LLM Context: Web search for AI agents and chatbots", undated
2. Brave Search API, "LLM Context: Web search API for AI agents", undated

---

Check how a page reads to answer engines. Open the GEO score analyzer